DeepSeek Unveils Three Major Moves in Single Week
Between September 8 and 10, 2026, DeepSeek simultaneously launched three critical updates: its new lightweight model V4.1-Flash on September 10, a historic 150-person hiring drive on September 8, and a commitment to continue V4 Pro API services beyond September 14. These concurrent moves span product, organizational, and commercial layers, marking the company’s most intensive adjustment cycle to date.
- Launch Date: September 10, 2026, 12:00 (Beijing Time)
- New Model: DeepSeek-V4.1-Flash, 552B-parameter MoE model with 8B input activation and 16B output activation
- Pricing Changes: Off-peak prices set at 0.02 RMB (cached hit), 1 RMB (uncached input), and 4 RMB (output) per 1K tokens; peak时段 pricing doubles
- Availability: Effective immediately as of September 10, 2026
- V4 Pro Continuation: API will remain available post-September 14, 2026, with unchanged billing terms
- Recruitment Scale: 150 roles (largest single recruitment), covering server-side development and Agent elastic computing
Product Evolution and Organizational Expansion
V4.1-Flash is branded as the smallest model in DeepSeek’s new structural series, featuring native multimodal visual understanding and outperforming flagship models like V4-Pro across multiple benchmarks. Its innovation lies in asymmetric input-output design—activating only 8B parameters for input and 16B for output—yielding notable cost savings over symmetric architectures.
Parallel to product innovation, the DeepSeek Harness team leader Cui Tianyi explained that the 150-person expansion addresses escalating operational complexity: “Dramatic increases in data volume, machine/container count, and training task load have rendered existing backend systems insufficient.” Positions span six areas: large model research platforms, Agent framework components, R&D efficiency infrastructure, DeepSeek API, online services, and data engineering, plus platform and systems development for Agent elastic computing.
A key counterintuitive fact emerges: Though labeled “smallest,” V4.1-Flash’s 552B MoE parameters dwarf many competitors, with true competitiveness rooted in structural optimization rather than parameter scale—its benchmark performance surpasses V4-Pro despite being technically lighter.
| Model Version | Parameter Size | Active Parameters | Multimodal Capability | Positioning |
|---|---|---|---|---|
| V4.1-Flash | 552B MoE | Input 8B / Output 16B | Native Support | Smallest size, intelligence exceeds V4-Pro |
| V4-Pro | Undisclosed | Undisclosed | Not specified | Flagship model, now outperformed by V4.1-Flash |
Service Continuation and Ecosystem Alignment
To ease user transition during model upgrades, DeepSeek confirmed V4 Pro’s continued availability: API access will persist beyond September 14, 2026, with unchanged billing. This avoids immediate reengineering pressure on downstream integrations, demonstrating commercial stability.
Simultaneously, DeepSeek operates within broader industry validation frameworks. Its models face vulnerability testing from entities like Ant Group’s AI security lab, which achieved 98.5% vulnerability verification success on CyberGym, alongside DeepSeek and others in benchmark evaluations.
Practical Recommendations
- Ready to adopt: Developers requiring cost-sensitive, lightweight multimodal reasoning (e.g., mobile agents, embedded inference) benefit from V4.1-Flash’s 16B output activation, dramatically reducing inference costs;
- Worth waiting: Teams deeply integrated with V4-Pro and complex job scheduling should monitor September 14 transition compatibility or evaluate migration effort to V4.1-Flash.
A Quiet Phase of Maturation
The combination of “open-source models + private deployment + V4 legacy support” reflects an industry entering a rational growth stage: while global giants delay IPOs amid security concerns, Chinese teams like DeepSeek prioritize cost optimization and commercial stability amid technical acceleration.
（Word count: 1497）